Research On Group Behavior Prediction Method Based On Generative Adversarial Mechanism

In modern society,predicting the group behavior in crowd scenes and obtaining the future trajectories of interaction between multiple crowds have become research hotspots in complex scenes,and have attracted extensive attention in the field of computer vision.Based on the existing trajectory prediction framework,combined the theoretical basis of each representative trajectory prediction algorithm,this paper proposes a new interpretable model based on generative adversarial networks.In this paper,the historical trajectory data of all pedestrians in the scene and the original scene picture data are used as the input of the model.Based on the generative adversarial network framework,a new physical attention mechanism and social attention mechanism are newly introduced to double consider the trajectory prediction.Pedestrians' future trajectories more in line with physical and social norms.The specific research contents of this article are as follows:(1)On the basis of generative adversarial network framework,a new physical attention mechanism and social attention mechanism are newly introduced.The' physical attention module helps the network to extract background information such as the terrain of the scene to obtain more important areas that are worthy of attention,avoid invalid predictions,and increase the accuracy of predictions;the social attention module can aggregate information between different pedestrians and extract information from different pedestrians.The most important trajectory information of pedestrians around makes this method predict a path that conforms to social norms.The supplement of the double attention mechanism of physics and society makes up for the problem that the traditional prediction algorithm does not take into account the environment information of the scene,and realizes a more reasonable and accurate prediction of pedestrians.(2)This paper uses the generative adversarial network as the basic framework of the prediction algorithm,so that this model can also achieve multi-modal prediction of the pedestrian's future trajectory.The essence of trajectory multimodality is that based on the historical information of pedestrian trajectories and scenes,pedestrians can have multiple acceptable future paths.Through testing on multiple simulation datasets,we verified that the model in this paper does possess multi-modal generation properties.Considering the prediction of the future trajectory of the crowd,we first cluster the trajectories of different crowds in the scene to obtain the clustering result of each crowd trajectory.Then input this clustering result as mean trajectory data to the above model,roughly predict the trajectory of its future moment,this is the trend direction prediction of the crowd.(3)This paper has conducted experimental tests on the real benchmark datasets ETH and UCY,and the data index results can prove that this model is superior to simple linear networks,long-short-term memory networks and derived social long-short-term memory networks and social generation confrontation network trajectory prediction models in terms of prediction accuracy.In terms of crowd trend prediction,we first verified the accuracy of the crowd clustering algorithm,then visualized the prediction of the crowd trend,and also obtained an accurate prediction of the trend direction of the crowd.